Applied Speech and Audio Processing: With matlab examples
Download 2.66 Mb. Pdf ko'rish
|
Applied Speech and Audio Processing With MATLAB Examples ( PDFDrive )
5.2. Parameterisation
103 h=[1, -0.9375]; % Apply the emphasis filter es=filter(h, 1, s); % Apply the de-emphasis filter ds=filter(1, h, es); It is instructive to try listening to this on a piece of test speech and to use the Matlab psd() function to conveniently plot a frequency response of the signals at each stage. The slightly nasal sounding es is a sound that many speech researchers are familiar with. If, when developing a speech algorithm, you hear something like this, then you will realise that you have lost some of the lower frequencies somewhere, or perhaps forgotten the de-emphasis filter. 5.2.2 Reflection coefficients Since the LPC coefficients by themselves cannot be reliably quantised without causing instability, significant research effort has gone into deriving more stable transformations of the parameters. The first major form are called reflection coefficients because they represent a model of the synthesis filter that, in physical terms, is a set of joined tubes of equal length but different diameter. In fact the same model will be used for Line Spectral Pairs (Section 5.2.4) but under slightly different conditions. The reflection coefficients quantify the energy reflected back from each join in the system. They are sometimes called partial correlation coeffi- cients or PARCOR after their method of derivation. Conversion between PARCOR and LPC is trivial, and in fact LPC coefficients are typically derived from input speech by way of a PARCOR analysis (although there are other methods). This method, and the rationale behind it, will be presented here, and can be followed in either [1] or [2]. First, we make the assumption that, given a frame of pseudo-stationary speech residual, the next sample at time instant n can be represented by a linear combination of the past P samples. This linear combination is given by Equation (5.5): x [n] = a 1 x [n − 1] + a 2 x [n − 2] + a 3 x [n − 3] + · · · + a P x [n − P]. (5.5) The error between the predicted sample and the actual next sample quantifies the ability of the system to predict accurately, and as such we need to minimise this: e [n] = x[n] − x [n]. (5.6) The optimum would be to minimise the mean-squared error over all n samples: E = n e 2 [n] = n x [n] − P k =1 a k x [n − k] 2 . (5.7) |
Ma'lumotlar bazasi mualliflik huquqi bilan himoyalangan ©fayllar.org 2024
ma'muriyatiga murojaat qiling
ma'muriyatiga murojaat qiling